An automated machine-learning model trained on preoperative CT angiography data correctly identified the primary entry tear location in Stanford Type A aortic dissection with an AUC of 0.862 internally and 0.785 in external validation. Built on 680 development patients and validated on 89 independent cases, the random forest classifier distilled 131 morphological features to just five — capturing ascending false lumen and thrombus burden, arch geometry, and ascending aortic caliber — achieving 97.6% sensitivity at a pre-locked threshold in external testing, though specificity collapsed to 17%.
Locating the primary entry tear before emergency surgery matters enormously: it directly dictates how far surgeons must extend the aortic repair, and misclassification can mean either under-resection or unnecessary operative risk. Current emergency CTA interpretation is subjective and time-pressured, making an automated assist clinically appealing. However, the external validation specificity of 17% signals a serious threshold-transportability problem — a well-documented challenge when models trained on one institution's case mix meet different demographics or imaging protocols. The exploratory cohort-specific threshold recovered reasonable balance (78.6%/68.1%), confirming the classifier has genuine signal but that thresholds must be locally calibrated. The retrospective single-center development design and modest external cohort (89 patients) further limit generalizability. As a preprint not yet peer-reviewed, these findings could change materially. Considered incrementally promising rather than practice-changing, this work sets a credible foundation for prospective multicenter trials before any clinical deployment.